agenthours.dev

LOCAL CODEX TELEMETRY / OPEN METHOD

AGENT HOUR TRACKER 30-DAY WINDOW

How many hours a day does your Codex work?

Several Codex agents can work while your editor feels quiet. Agent Hour Tracker counts completed parallel turns in your local archive.

START WITH A SANITIZED SCORE

uvx codex-agent-hour-tracker --share

Runs against the local Codex session archive. The command needs no account or upload.

Open the GitHub repository

01 / WHAT THE NUMBER MEANS

Cumulative time, not wall-clock time.

Cumulative agent-hours add the completed duration of each agent turn. The tracker sums overlapping turns independently: three agents running for an hour count as three agent-hours, even though only one hour passed on the clock.

ARCHIVE SCORE

Mean cumulative agent-hours per calendar day across exactly the 30 most recent completed local calendar days ending yesterday, including zero-use days.

02 / RUN IT LOCALLY

One command. An inspectable result.

Use uvx to run the published CLI without a project checkout. It scans completed root and delegated turns, applies bounded timing rules, and prints the sanitized aggregate card.

  1. 01

    Paste the command above into a terminal with your local Codex archive.

  2. 02

    Read the score, dates, and diagnostics before deciding whether to share.

  3. 03

    Open the published package or inspect the methodology when you want the rules.

03 / PRIVACY BOUNDARY

Measure the archive without opening the conversation.

Calculation stays local. The scanner retains bounded timing metadata and does not decode conversation content. There is no upload through this site or the CLI.

Share card: --share is a bounded aggregate designed for sharing, but it is still an intentional disclosure.

Detailed reports: Full and CSV reports reveal semi-sensitive day-level patterns. Review them before distribution.

Source: The page has no analytics, cookies, external scripts, or data connection.

04 / METHODOLOGY

A small set of explicit rules.

The tracker attributes each completed turn to its local start date. It caps event gaps at 30 minutes, uses bounded fallbacks for sparse timelines, and removes duplicate turn IDs before aggregation. The full rules and limitations are public.